Abstract: The construction industry, particularly in civil engineering and oil reservoir protection earthworks, continues to grapple with stagnant equipment productivity and rising pressure to decarbonize operations, yet traditional planning and reactive management approaches remain unable to leverage the high-frequency data now streaming from modern heavy machinery. This study addresses the critical gap between the growing digitization of construction assets and the lack of an integrated framework that converts real-time sensor data into predictive, multi-objective operational decisions for earthmoving tasks vital to oil reservoir safety infrastructure. A closed-loop decision framework is proposed that seamlessly fuses a synchronized equipment digital twin with a suite of machine learning models and a receding-horizon optimizer to simultaneously maximize earthmoving fleet productivity and minimize carbon emissions. The research was conducted on data from a four-week medium-scale earthmoving operation involving the construction of protective containment berms around oil storage reservoirs, employing two 20‑ton excavators, five articulated dump trucks, and a wheel loader, all instrumented with CAN bus loggers, hydraulic pressure sensors, inertial measurement units, and GNSS receivers generating over five terabytes of high-frequency time-series data. Digital twins for each machine were instantiated using multi-domain physics-based models, calibrated to field data, and continuously synchronized via an inverse dynamic’s payload estimation algorithm and a state detection engine. A total of 137 engineered features were extracted per work cycle, spanning temporal statistics, spectral signatures, and domain-specific ratios such as idle time ratio and hydraulic load factor, forming a 14,000‑cycle feature matrix. Five distinct machine learning architectures—extreme gradient boosting, random forest, deep neural network, long short-term memory network, and Transformer—were trained to predict productivity rate and fuel consumption, with log-transformed targets and strict time-series cross-validation. The extreme gradient boosting model achieved a coefficient of determination of 0.893 for productivity prediction on the hold-out test set, outperforming a static baseline (0.317) and a daily-aggregate naive model (0.612) by wide margins, while the long short-term memory network led fuel consumption prediction with an R² of 0.904. Feature importance analysis revealed that effective payload, idle time ratio, and hydraulic load factor accounted for over forty percent of predictive power, providing an actionable prioritization for instrumentation investments. The trained models were embedded as surrogate evaluators within a multi-objective Chebyshev-scalarized genetic algorithm that re-solved every minute over a ten-minute horizon, guiding throttle settings, bucket fill depth, and truck dispatch. In a 60-day simulated project, the integrated optimizer under equal productivity-emission weights moved 108,420 cubic meters of material and consumed 41,200 liters of fuel, simultaneously surpassing reactive dispatch by 9.8% in output and 4.4% in fuel efficiency. The Pareto frontier mapped across five weight configurations exposed a concave trade-off with a marginal emission cost of 1.98 kg CO₂ per additional cubic meter at the productivity extreme and a marginal sacrifice of 0.90 kg CO₂ saved per lost cubic meter at the emission extreme, a decision-support instrument without precedent in the literature. Robustness experiments with artificial sensor dropout rates up to fifty percent demonstrated only a 12.8% relative degradation in predictive accuracy for the gradient boosting model, and the optimizer maintained a cumulative output of 104,100 cubic meters under thirty percent dropout, still exceeding the full-data reactive strategy. These findings confirm all three embedded hypotheses: real-time digital twin data significantly enhances prediction accuracy; multi-objective optimization identifies superior productivity–emission trade-offs; and the integrated system exhibits field-grade resilience to data imperfections. The study provides a validated, computationally tractable blueprint for anticipatory equipment management that is ready for field piloting and scale-up in safety-critical earthworks such as oil reservoir containment berms, marking a decisive shift from descriptive monitoring toward predictive, self-optimizing construction operations.
Dizaji* et al. (Wed,) studied this question.